I built NeuroSync originally as a final year school project to explore Neural Cryptography. The main idea is the use of adversarial learning to train two NNs (Alice and Bob) to encrypt and decrypt data while a third one (Eve) is trying to break the cipher. While that isn't new I made it into a usable library with an interface that allows to easily switch parameters and experiment with them to help research in this field.

To make it functional, I implemented a hash-based correction system to ensure 100% decryption accuracy (bypassing the usual 0.2% probabilistic errors of neural synchronization), alongside packet sequencing and periodic key changes.

It is definitely still a work in progress, and the "Eve" (attacker) network currently needs a lot of work. But the first version is usable and the core communication protocol works.

I added a Jupyter Notebook in the repo so you can test the key exchange and encryption without setting anything up. I'm sharing this mostly to get your feedback, hear your ideas, and maybe find some folks interested in this field too.

Would love to hear your thoughts or critiques!